Skip to content

v0.9.0: Granular Specialist Skills Architecture

Choose a tag to compare

View all tags
@githubmofo githubmofo released this 02 Sep 17:48
· 12 commits to main since this release
v0.9.0
38d676f

TorusGuard v0.9.0: Granular Specialist Skills Architecture

Release Version: v0.9.0
Date: September 2, 2026
Status: Official Release

馃殌 Overview

TorusGuard v0.9.0 introduces a modular, decoupled specialist skills architecture. Rather than loading a monolithic, context-heavy instruction set on every prompt, TorusGuard decomposes all security operations into 12 self-contained specialist skills plus an intelligent lazy-loading router.

Each specialist skill embeds its own execution steps, safety rules, and scoring models inline, strictly constrained within a 300-line context budget (ranging from 58 to 165 lines).

馃専 Key Highlights

1. The 12 Specialist Skills (skills/)

  • skills/torusguard-init/: Baseline discovery, stack detection, and tailored rule activation.
  • skills/torusguard-authorize/: Scope boundary capture, target ownership verification, and safety policy enforcement.
  • skills/torusguard-audit/: Deep static AST analysis, invariant line-shift fingerprinting, and root-cause clustering.
  • skills/torusguard-verify/: Evidence sufficiency evaluation and 5-factor confidence score refinement.
  • skills/torusguard-web-validate/: Authorized HTTP probing, session tracking, and automatic credential redaction.
  • skills/torusguard-exploit-check/: Bounded, non-destructive canaries confirming exploitability across 5 canonical statuses.
  • skills/torusguard-harden/: Ponytail Protocol patch formulation ($\le 35$ additions, $\le 25$ deletions) and bundle packaging.
  • skills/torusguard-apply/: Surgical patch application with mandatory pre-apply rollback snapshots.
  • skills/torusguard-recheck/: Targeted post-patch re-scans with 4-state transition tracking.
  • skills/torusguard-report/: Executive markdown reporting and OASIS SARIF v2.1.0 structured export.
  • skills/torusguard-status/: Read-only workspace inspection of active configuration and run history.
  • skills/torusguard-full/: Master 7-stage pipeline orchestrator coordinating end-to-end security operations.

2. Intelligent Lazy-Loading Router (skills/torusguard/SKILL.md)

The root torusguard skill acts as an ultra-compact (58 lines) router. When an operator runs a slash command (e.g. /torusguard audit), the AI agent loads exclusively the matching specialist file, eliminating cross-file context bloat and leaving 98%+ of the context window free for source code analysis.

3. Automated Validation Harness (harness/validate_v0_9_0_skills.py)

53 automated checks validating skill existence, YAML frontmatter, line budgets ($\le 300$), required sections, router table integrity, and script bindings.

Assets 2
Loading